Proxmox VE vs OpenNebula vs OpenStack¶
Canonical comparison of the three open-source private-infrastructure platforms in this knowledge base. They are frequently weighed against each other but occupy different ambition levels; the practical question is not "which is best" but "how much cloud do I actually need". Topic folders: Proxmox VE · OpenNebula · OpenStack. For container-centric stacks, see the sibling Kubernetes vs OpenNebula vs OpenStack comparison.
TL;DR¶
| Proxmox VE | OpenNebula | OpenStack | |
|---|---|---|---|
| What it is | Hypervisor appliance: Debian + KVM/LXC + cluster stack in one product | Cloud manager: centralized controller (oned) orchestrating KVM/LXC/vGPU fleets |
Cloud OS: 50+ cooperating services for multi-tenant IaaS |
| Design center | Per-node simplicity; a cluster of 3-50 nodes as one unit | One control plane over many hypervisors, incl. edge | Massive scale-out, many teams, many tenants |
| Tenancy | Teams/RBAC on one admin plane | Groups/VDNs, self-service portal, hybrid federation | Full multi-tenant: quotas, projects, billing-ready |
| Scale sweet spot | 1-50 nodes | 10s-100s of nodes, edge fan-out | 100s-1000s of nodes |
| Current version | 9.2 (May 2026; arm64 Aug 2026) | 7.2 (Apr 2026) | 2026.1 "Gazpacho" (Apr 2026, SLURP) |
| Storage story | Ceph + ZFS built in; PBS backup integrated | Datastore abstraction (image/system/data backends over Ceph, NFS, or local FS) — bring-your-own philosophy | Cinder/Ceph/Glance: pluggable, enterprise-grade, many backends |
| Backup | PBS (dedup, encryption, live restore) part of the family | Image + datastore-centric; BYO backup | BYO (Freezer/third-party) |
| License / cost | AGPLv3; EUR 120-1,100/socket/yr support tiers | Apache 2.0; enterprise subscription (quote-based) | Apache 2.0; free upstream, paid via distros (Canonical/Red Hat/Mirantis) or self-support |
| Ops skill floor | One admin can run it | One small team can run it | A platform team is expected |
| Best for | VMware replacement, homelab→enterprise clusters, hyperconverged branches | Sovereign/hybrid/edge clouds, AI/GPU factories, MSPs | Multi-tenant IaaS, large enterprises, public-cloud-style internal platforms |
Layer Mapping¶
The confusion dissolves once the three are placed on an ambition axis rather than treated as peers:
flowchart TB
subgraph L1["Layer 1 - hypervisor appliance (per-node product)"]
PVE["Proxmox VE<br/>KVM + LXC + pmxcfs cluster + CRS<br/>Ceph/ZFS + PBS + SDN built in"]
end
subgraph L2["Layer 2 - cloud manager (one control plane)"]
ONE["OpenNebula<br/>oned central controller<br/>Sunstone portal, gRPC API (7.2)<br/>edge + AI factory focus"]
end
subgraph L3["Layer 3 - cloud operating system"]
OS["OpenStack<br/>Nova/Neutron/Cinder/Glance/Keystone...<br/>message bus + SQL per service<br/>SLURP release train"]
end
HW["Commodity servers (KVM hypervisors)"]
PVE --> HW
ONE --> HW
OS --> HW
PVE -. "grows into: PDM, Cattle-and-Pets roadmap" .-> L2
ONE -. "grows into: federation at scale" .-> L3
All three run the same workloads at the bottom (KVM guests, Linux containers, Ceph). The difference is how much cloud — tenancy, self-service, quota, multi-team operations — wraps around them.
Architecture Comparison¶
Proxmox VE¶
The stack lives on every node: web UI/API (pveproxy), daemons, storage plugins, and the corosync-replicated cluster filesystem (/etc/pve, 128 MiB cap). There is no external controller to size or babysit — the cluster is the product. Config replication gives duplicate-ID protection and quorum-gated safety; HA is watchdog fencing plus the 9.2 CRS dynamic load balancer. Trade: a bounded management plane (thousands of guests, not tens of thousands) and corosync network discipline.
OpenNebula¶
Deliberately centralized: one oned daemon (with gRPC API since 7.2) holds the state, fronted by the Sunstone portal and Flow/VOne services for multi-VM apps. Failover of the controller is via replicated front-ends (RAID + DB HA), not a distributed consensus layer — drastically simpler to reason about and troubleshoot than OpenStack's service mesh of queues. Hypervisors stay thin (KVM/LXC + monitoring agents). Strengths: production in hours, native NVIDIA GPU/vGPU orchestration aimed at AI factories, edge fan-out (lightweight remote clusters syncing to one control plane), and VMware import tooling.
OpenStack¶
A constellation: Nova (compute), Neutron (network), Cinder (block), Glance (images), Keystone (identity), Horizon (UI) — each an independent service with its own database, communicating over a message bus. That architecture is what buys true multi-tenancy, quotas, per-service HA, and an unmatched integration catalog — and what demands a platform team, release-train discipline (SLURP releases like 2026.1 Gazpacho, direct upgrades from 2025.1 Epoxy), and deliberate capacity engineering.
Operations Compared¶
| Dimension | Proxmox VE | OpenNebula | OpenStack |
|---|---|---|---|
| Deploy | ISO per node, join cluster (hours) | Appliance/ansible to front-end + hosts (hours-day) | Deployment tooling (Kolla/Ansible/Helm) — days-weeks |
| Upgrade | apt + major-version guide (8→9 seamless) | Controller-centric upgrades | Release-train SLURP upgrades; significant planning |
| Day-2 backup | PBS integrated (dedup, live restore) | BYO (image/datastore exports) | BYO / third-party |
| GUI | Integrated per cluster | Sunstone cloud portal | Horizon (dense, less polished) |
| Skills | Linux + virtualization admin | Same, lighter | Linux + Python services + messaging + DBs + networking depth |
| Break-glass | Everything is plain text under /etc/pve | Centralized DB + config — one place to look | Per-service debugging across the constellation |
Networking and Tenancy¶
- PVE: Linux bridges/VLANs plus its SDN (zones: Simple/VLAN/QinQ/EVPN; 9.2 fabrics — WireGuard/BGP, route maps, IPv6 underlay). Tenancy = RBAC + pools within one admin plane.
- OpenNebula: VNets with VLAN/VXLAN isolation, security groups, route tables; multi-tenancy via groups + VDCs (virtual datacenters), plus hybrid cloud connections (AWS/Azure bursting) and edge cluster federation.
- OpenStack: Neutron is the deepest of the three — full overlay networking (VXLAN/Geneve, OVN), LBaaS, FWaaS, QoS, arbitrary tenant routers — the price of running a network OS inside your cloud.
Decision Matrix¶
| Dimension | Proxmox VE | OpenNebula | OpenStack |
|---|---|---|---|
| Architecture | Monolithic per-node cluster stack | Centralized controller + thin nodes | Distributed service constellation |
| Performance | Excellent (near-bare-metal KVM; CRS balancing) | Excellent (thin hypervisor path) | Good; more virtualization layers per packet/request |
| Operations | Simplest (one product) | Simple (one controller) | Most complex (platform team) |
| Cost | Free; EUR 120-1,100/socket/yr support | Free; subscription for enterprise features/support | Free upstream; distro support or self-support economics |
| Security | Tight surface (mgmt plane = product; corosync isolation; 2FA/RBAC) | Controller is the crown jewel; standard hardening | Keystone RBAC + per-service policy — strong, but wide surface |
| Ecosystem | PBS, PDM, TF/Ansible providers, big community | GPU/edge/hybrid focus, gRPC, VMware migration | Largest: drivers, vendor distros, public-cloud parity features |
| Scale ceiling | ~50-node clusters (pmxcfs/CRS bound) | Hundreds incl. edge | Thousands of nodes, many tenants |
| Lock-in risk | Low (plain-text config, open API, portable images) | Low (Apache 2.0, open formats) | Low per se; distro-version gravity |
Weighted Decision Matrix¶
Scenario A: ** Replace VMware for an internal platform of 5-20 nodes, one infra team, mixed VMs + some GPU.** Scale 1-5, weights sum 100%.
| Criterion | Weight | PVE | OpenNebula | OpenStack |
|---|---|---|---|---|
| Operational simplicity | 25% | 5 | 4 | 2 |
| Time to production | 15% | 5 | 4 | 2 |
| Backup/DR story | 10% | 5 | 3 | 3 |
| Cost (license + support) | 10% | 5 | 4 | 3 |
| Self-service/tenancy needs | 10% | 2 | 4 | 5 |
| Scale headroom | 10% | 3 | 4 | 5 |
| Ecosystem/talent | 10% | 4 | 3 | 4 |
| GPU/AI features | 10% | 4 (passthrough, NVIDIA vGPU) | 5 (native orchestration) | 3 |
| Weighted total (max 5.00) | 4.25 | 3.85 | 2.95 |
Scenario B flips the self-service/scale weights (multi-tenant, 200+ nodes, multiple business units): OpenStack 4.05, OpenNebula 3.80, PVE 3.05. The matrices are mirror images — which is the point: these platforms win at different altitudes.
Verdict¶
- Primary — Proxmox VE when the requirement is a virtualization cluster, not a cloud: one team, tens of nodes, want enterprise backup in the box, and VMware-exit pressure. It is the fastest path from decision to production and the lowest operational floor. Escape to a "real cloud" layer later is non-destructive — images port, PBS restores anywhere.
- Choose OpenNebula when those nodes must behave like a cloud — self-service portal, groups/VDC tenancy, edge sites fanning out, GPU factories with native NVIDIA orchestration — but a platform team for a full constellation is not on the menu. It is the deliberate middle: 10x simpler than OpenStack, considerably more cloud than a hypervisor cluster.
- Choose OpenStack when multiple business units need isolated tenants with quotas and self-service at serious scale, when public-cloud-parity networking (OVN, LBaaS, FWaaS) is a requirement, or when a vendor distro (Canonical/Red Hat/Mirantis) is acceptable to carry the operational load.
- Sequencing note: these are stages, not rivals — common and defensible path is PVE now, OpenNebula (or OpenStack via a distro) when self-service demand arrives.
Sources¶
All URLs verified HTTP 200 on 2026-09-25 unless noted. Facts about OpenNebula and OpenStack align with this vault's own topic notes (opennebula/index.md, openstack/index.md — checked 2026-09-25).
- Proxmox VE and pricing — official product and subscription pages
- PVE Roadmap — 9.2, CRS, Cattle-and-Pets
- OpenNebula — current release and features (v7.2)
- OpenStack releases — 2026.1 Gazpacho SLURP train
- Infrastructure Platforms Comparison (K8s vs OpenNebula vs OpenStack) — sibling comparison in this folder
- Vault topic notes: Proxmox · OpenNebula · OpenStack — per-topic sources therein